Follow the evidence

Follow agent activity, useful work, paid infrastructure use and market expectations. Prices and feeds update each trading day, filings each quarter and private rounds monthly. The capability chart retains METR’s historical Time Horizon 1.1 series. The baskets are hypothetical; this is research commentary, not investment advice. See the full disclosures in Appendix D3.

The companies

The indices

Hypothetical equal-weight baskets defined in 2026 and applied back to 2022; hindsight-selected, no fees, taxes or trading costs, not an investable index.

The layers of the stack · equal-weight indices · change since the base date

One line per layer of Figure 3.1, with State and Work drawn together as the backend, members as the Appendix C1 roster stands today. Equal weight, reset daily, chained from January 3, 2022, each company’s return in its own listing currency. The backend line is the published 32-name index; the other four layer lines follow the roster as it changes and are context. The Nasdaq 100 is the benchmark, rebased to the same day. The axis is the change since the chosen start on a log scale, so equal distances are equal percentage moves; the buttons above rebase every line, and clicking a legend entry hides that layer. This is Figure 4.1 in Part Four. Intelligence has two members, both listed in 2026, so its line starts at 0% on its first trading day, and a Hong Kong holiday carries the prior level.

The five-layer stack of machine labor

The October 1, 2026 map groups public companies by the services they provide. State and Work form the backend research universe. The other layers provide context; hyperscalers and SpaceX span several layers. Watchlist companies are listed but excluded from the count. The full roster is in Appendix C1. This is Figure 3.1 in Part Three.

Fundamentals

Valuation

as of the latest close · forward figures use the latest consensus

Enterprise value is market value plus debt less cash and short-term investments.

The macro signals

These are the series we read, not numbers we forecast. Tokens: how much of one routing platform is agents at work, what they do, what it costs, and the machine hours that implies. Other adoption evidence: the outside series behind the model’s inputs. Capability: how hard a task a machine completes, measured in the human time it would take (METR), not how long an agent runs. Production: how much software gets written. We grade the bold bets of Part Four against them.

Tokens

Estimated machine hours worldwide

Three assumptions determine the estimate. Their uncertainty is not combined into a confidence interval.

No worldwide count exists. We estimate the same quantity in two ways. The provider-by-provider estimate combines platforms: daily tokens × estimated agent share ÷ tokens per agent hour. We remove routed traffic from the provider totals so it is counted once. It updates as new disclosures arrive. Workload assumptions are in Appendix B1. The OpenRouter-based estimate starts with OpenRouter’s agent tokens. We convert coding, research and workflow tokens to hours, then divide by OpenRouter’s estimated share of worldwide token use. Unclear task tags follow a stated allocation rule, with alternatives shown. This method assumes one platform’s mix helps describe the wider market. The shading varies token use per hour across the three workload profiles. It does not include all uncertainty about worldwide token volume, classification or whether OpenRouter’s mix represents the wider market. Activity converted to hours, shown in the reading, provides another comparison from products that bill hours or actions. It is not a measured lower bound. Both main estimates share workload assumptions, so agreement is not independent confirmation and their range is not a confidence interval. The dashed lines show the published forecast: a fixed starting point of 35 million hours a day, set on October 6, 2026, and each scenario’s path through 2027. The marked method changes are explained in Appendix A4. Estimates may miss uncovered platforms and tool work that uses few tokens.

Token use on OpenRouter · agent, chat and mixed apps · trillion a day · 7-day average

Total platform traffic, split by application type. Dashed history through August 7, 2026 was read from OpenRouter’s published chart; later dots use weekly volumes and app classification. Lines between dots are guides, not daily observations. The marked method change matters when comparing periods. This is one platform’s traffic.

Estimated agent hours on OpenRouter · million hours a day · assumption range shaded

OpenRouter against the world’s tokens · by player, trillion tokens a day · quarterly, 2025 to the latest quarter

Both views interpolate between dated statements and hold the last value until a new disclosure. The default view includes estimated providers; the other view includes disclosed providers only. Quarterly bars are estimates at quarter-end, not direct measurements of every date. Coverage is incomplete: several major providers publish no volume, and OpenAI’s API disclosure excludes consumer ChatGPT. Our estimates fill gaps where providers publish no volume and where the China residual cannot be reconciled from the disclosed inputs. Estimated values appear as outlines and begin in the stated quarters. OpenRouter’s share changes with the selected total. The source record and calculation are under Method.

Assumptions · OpenRouter’s global share and workload mix

Worldwide hours equal estimated platform hours divided by the platform’s global share. Halving that share doubles the result. The starting share includes our estimates for providers without disclosures. OpenRouter classifies sampled traffic by task; we allocate those tasks to workload profiles. The two breakdowns use different denominators.

What an active agent hour uses · processed, fresh input and output tokens

Our session logs count intervals when the model generates or a tool runs, capped at ten minutes. Waiting for a person is excluded; a parent’s wait for a child counts only in the child’s work. Processed tokens include everything read and written. Fresh input includes new input and cache writes; output is reported separately. This is one operator’s sample. Appendix A2 preserves the earlier capped-gap estimate.

Tokens per request on OpenRouter · agent apps · trailing 30 days

OpenRouter publishes tokens and requests by app over the preceding thirty days. Dividing them estimates tokens processed per request. Comparing this with fresh input per call in our logs helps estimate how much context is being reused. App classification is in Appendix A4; the limits of that comparison are in Appendix A2.

What the tokens are used for · task shares · 7-day windows

OpenRouter’s task labels are grouped into agent work, code, data and general use. Each point covers seven days. The record begins September 27, 2026. This task-based definition of agent work is narrower than the app-based classification above, so the shares differ.

What tokens cost

Price of a million tokens · what the busiest models charge, weighted by use · dollars · since June 2025

Input is what the model reads; output is what it writes. The chart weights OpenRouter list prices by model token use. Weekly observations start in June 2025, with daily observations from September 28, 2026 covering the twenty busiest paid models. The join changes the basket as well as the frequency; it should not be read as a pure price change.

What a Claude Code session costs

A session is one run; a turn is one exchange with the model. The lines group sessions by turn count and weight OpenRouter’s median costs across the models used. Each point is a daily snapshot.

Other adoption evidence

These sources help assess the model’s inputs: business and employee adoption, spending, autonomy, agent activity, tool use and web traffic. Each observation retains its source and date. Comparable readings form charts; one-off disclosures remain in the tables in Appendix B4, and Appendix B1 says which input each series grades. Refreshed monthly, Cloudflare Radar weekly.

Capability

Task length the best public model completes · human-expert minutes · log scale

Historical METR Time Horizon 1.1 series. The source page is no longer actively updated (notice dated September 8, 2026). Model release dates are shown; is the fetch date, not a new evaluation date. METR expresses task difficulty as the time a human expert would need. The chart shows frontier models at 50% success on a logarithmic scale. It does not measure how long the agent ran. Estimates above sixteen expert hours are excluded because the task suite cannot support them. The table includes each evaluated model and its confidence interval.

Production

Merged pull requests a month

The share merged by coding agents is in Appendix B3.

New repositories a month

Public repositories only; the running month is never drawn. The weaker commits series is in Appendix B3.

The predictions

Dated predictions · measures, deadlines and results · launch bets, October 8, 2026

Each prediction has a measure and a deadline. We grade it against the named source and retain the original terms. The same record appears in summary Figure 0.4 and Part Four’s Figure 4.8.

Research commentary from Skycatcher, LLC, an exempt reporting adviser, not investment advice. The baskets are hypothetical, hindsight-selected and not investable; the scenario figures are our estimates, not company guidance. Sources: SEC EDGAR, Financial Modeling Prep, METR, the GitHub Search API and OpenRouter. Nasdaq-100 is a registered trademark of Nasdaq, Inc., which does not sponsor or endorse this site. Full disclosures in Appendix D3.

For the dated reading 2026-10-25, the OpenRouter-based estimate is 53 million machine hours a day worldwide. The provider-by-provider estimate is 17 million. Neither is a direct count; both depend on workload assumptions.